A Feature Selection Approach to Visual Domain Adaptation in Classification

摘要

In machine learning, we presume datasets to be labeled while performing any operation. But, is it true in real-life scenarios? To its contrary, we have an enormous amount of unlabeled datasets available in the form of images, videos, audios, articles, and many more. The major challenge we face is to train our classification model with primitive machine learning algorithms because these algorithms only expect labeled data. To overcome these limitations visual domain adaptation algorithms such as MEDA (Manifold Embedded Distribution Alignment) have been introduced. The main motto of MEDA is to minimize the distribution difference between the source domain (an application that contains enough labeled data) and the target domain (an application that contains only unlabeled data). In this way, the source domain labeled data can be utilized to enhance the performance of the target domain classifier. Though MEDA (Manifold Embedded Distribution Alignment) approach shows remarkable improvement in classification accuracy, but still there is considerable scope of improvement. There are plenty of irrelevant features in both domains. These irrelevant features create a hole for this algorithm and prevent the target domain classifier from becoming more robust. Therefore, for the purpose of filling this hole, we propose a new feature selection based visual domain adaptation (FSVDA) method which uses particle swarm optimization (PSO), where the MEDA method is considered as a fitness function that leads to automatically select a good subset of features over both the domains. Extensive experimental results on two real-world domain adaptation (DA) data sets such as object recognition and digit recognition demonstrate that our proposed method outperforms state-of-the-art primitive and domain DA algorithms. It is a big challenge to train the classifier for a new unlabeled image dataset in image classification and computer vision. The two magnificent solutions to this challenge are transfer learning and domain adaptation. By transfer learning, we can use our knowledge from previously trained models for training newer models.